TouchScale: 500-Hour Vision-Tactile Dataset Lifts Robot Manipulation Success from 22.5% to 57.5%
2077AIDataFoundation · hf · 2026-10-12
A multi-institution team (Texas A&M, Google DeepMind, CMU, Stanford and more) releases TouchScale, the largest vision-tactile dataset collected with a single unified wearable setup.
- Scale: 500 hours of contact-rich human interaction, 2K task descriptions, each synchronizing egocentric RGB-D video, wrist RGB, and dense full-hand bimanual tactile data.
- Perception: full-dataset training raises zero-shot contact IoU on an unseen tactile sensor from 0.134 to 0.383, and yields top action recognition accuracy on three benchmarks.
- Robotics: visual-tactile mid-training improves average real-world success on four contact-rich manipulation tasks from 22.5% to 57.5%.
- Scaling: with fixed sensing and protocol, both tactile prediction and robot success trend upward with more data. The dataset will be fully released.
More from Embodied
- Boston Dynamics' Atlas evolution 2019-2026: the gap is the story — CurieuxExplorer · 2026-10-12
- BYD's humanoid robot expands from factories to customer service — alexmacgregor__ · 2026-10-12
- Nature MI paper: microrobot navigation trained in just minutes — lulzxdxdxd · 2026-10-12
- Robotic systems and real-time connectivity let surgeons operate on patients thousands of km away — CurieuxExplorer · 2026-10-12
- Humanoid robots learn Nadal and Federer's tennis styles straight from broadcast footage — CurieuxExplorer · 2026-10-12
- Humanoid hands are the hardest problem: Shadow, Mimic and PSYONIC specialize in them — xmercury_one · 2026-10-12